#推理

Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights

Anthropic's Frontier Red Team released FlashProve, an open-source zkSNARK prototype that proves a reply came from a committed model without revealing its weights, reaching about 110 tokens/second on Llama 4 Maverick (400B MoE) with one H100 and 18.78 MB proofs verified on a CPU in 17 seconds. Proving costs roughly 70 to 95x the GPU time of plain vLLM inference, and the system is currently limited to 1,024-token contexts, greedy decoding and MoE models. Security properties are formalized in Lean, with a cheating provider caught except with probability about q/2^124 + 2^-101.

Fireworks unveils new products at Forge to help companies build specialized intelligence

At its first annual Forge conference, Fireworks launched products aimed at helping companies build AI around their own data and tasks, including Ultra-fast inference on NVIDIA Blackwell GPUs reaching up to 1,500 tokens per second, a half-price Background Mode, and multi-day prompt caching. It also introduced Fireworks Nexus for bringing open models into existing coding and coworking tools, No Eval and Specialized Model Inference for training and serving custom models, and Fireworks Sovereign, now in private preview, which runs its inference engine inside customers' own cloud or data centers. Fireworks says it already manages over 25,000 GPUs in customer environments.

Rapid LLM development is outrunning the research meant to explain it, study finds

Stanford researchers Owen Queen and James Zou tested 50 widely cited claims about large language models against 20 models and found 44 percent have already "expired," with an average lifespan of 15.7 months among the 22 findings with an estimated expiration date. Claims about what models can't do expired fastest, while findings on preferences, sycophancy and evaluator behavior proved longer-lived, and results varied widely across models even within the same release window. The researchers argue empirical LLM findings should be presented as observations about specific models at a specific point, with model, version, date, prompts and analysis code documented for re-testing.

Would Local AI Become the Dr.Pepper?

Using Dr Pepper's rise to America's No. 2 soda as an analogy, the post argues that local AI can earn a lasting place by fitting specific needs rather than outperforming frontier models. It notes limits of scaling—finite public training data and diminishing returns from added compute—and suggests tasks like routing, personal knowledge management, and dictation could move to edge devices where privacy, speed, and local control matter. It cites Google AI Edge Foresight, Fluid voice, LiteLLM, Ollama, and DSH's creator mode as examples for experimenting with locally hosted models.

I use AI (Oct 2026)

A research engineer at Salesforce AI Research reflects on how his AI use changed over 12 months, now managing roughly 5–6 coding agents per workday and spending about $3,000/month on tokens. He finds agentic coding fast and useful but still dependent on human taste, reports that LLMs have become his first choice for literature searches while remaining unreliable for novel research, and criticizes AI writing as mediocre and AI image and video generation as a net negative.

Cost per task: Using cheap classifiers to route across Gemini models and thinking levels

The article argues that cost per successful task, not price per token, governs production economics for coding agents, and reports a 50-question auto-graded benchmark routing tasks with TypeSafe's Jev classifier across Gemini configurations. A frozen regex and prompt-length heuristic scored 93.8% on 16 development questions but fell 23.2 points to 70.6% on 34 held-out tasks, while Jev agreed with hand labels on all 34 and 48 of 50 overall; capping maxOutputTokens at 300 saved nothing on easy tasks and broke 7 of 25 hard ones. A router sending tasks to Flash-Lite first and escalating Jev-flagged complex ones to Flash on low thinking solved all 50, as did Flash on low thinking alone, and the article notes caching favors keeping long context-heavy turns on one model.

StateBench: Testing how language models track state changes over long documents

Ai2 introduces StateBench, a long-context state-tracking benchmark with four difficulty levels (r-trivial, aperiodic, soluble, and general finite) grounded in Krohn–Rhodes theory. Testing open pretrained models and frontier models including Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna, they found accuracy fell sharply on soluble and general finite problems as inputs grew longer, dropping to 13% or lower on the longest general finite problems. In from-scratch training of 275-million-parameter models, Gated DeltaNet and hybrid models generalized almost perfectly on the first three levels out to tens of thousands of assignments, while transformers never learned the hardest level.

Qwen 3.8 27B on my laptop can interpret model weights, too

Simon Willison tested the "Sparks of Weight-Space Programming" paper's game-identification prompts, which ask a model to name which of four Atari-style games a policy network was trained on using only its raw weights and no tools. Running Qwen 3.8 27B locally via LM Studio, he got correct answers for Pac-Man (46m 12s, 43,506 reasoning tokens), Pong (1h 4m 54s) and Breakout (52m 25s), but it failed on Freeway. He notes critics like Matthew Honnibal and Sean Goedecke dispute the paper's framing, and he focuses on the no-tools question rather than claims that source code is obsolete.

Introducing Sarvam Prism

Sarvam released Sarvam Prism, a System One decision model built as a LoRA fine-tune of Sarvam 105B that returns a probability for each option in a single forward pass. Sarvam reports Prism scores 71.8 averaged over three indices at a 44 ms median response time on one B200 GPU, the highest-scoring model under 100 ms among the 25 it tested, and that its errors stay within its confidence bounds in English and Indian languages. Weights are on Hugging Face under Apache 2.0, and the model is also available on the Sarvam API via Sarvam Indus.

[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop

AINews reports a paper in which a smart model directly reads and edits the weights of a smaller model, described as one-shot fixing bugs from failed traces, improving in a general direction over about 50 steps, and writing entire network weights from scratch that beat some independently trained models. The post frames this as a new recursive-self-improvement capability that emerged suddenly, notes an independent reproduction appears to have already found it, and cautions it is unlikely to scale beyond simple games and RL policies for now. The same issue also covers DeepSeek's open-sourced "super-agent" harness prototype, the GDPx benchmark where the best model scores 12%, and October model releases.

Zyphra AI Introduces Zyphra Hierarchical Attention (ZHA): A Path to True Subquadratic Attention

Zyphra released Zyphra Hierarchical Attention (ZHA), a sparse attention method that uses a tree-structured indexer to search the KV cache instead of scanning the full context, and its paper shows it can be distilled into frozen pretrained models with about 4B tokens. On Qwen3.6-35B-A3B, ZHA adds 81,920 trainable parameters and stays within a few points of dense attention on seven reasoning, coding and retrieval benchmarks, with the wide tree reaching 96.2% on RULER-13 at 256K versus 96.0% dense, while MSA fell to 77.1%. Zyphra released no optimized kernels, so speedups such as 373 to 417x over dense at 1B tokens come from an H100 BF16 roofline model rather than measurements.

Ask HN: "Net neutrality" for AI compute and memory?

An Ask HN post asks whether something like "net neutrality" should apply to DRAM and compute, noting that both Sam Altman and Elon Musk have raised compute/DRAM scarcity and questioned whether local open weights models are the most sustainable solution. It asks whether allocation should follow a dynamic market-based price mechanism, go to the highest and best use, or let whoever pays use it for any purpose, and whether laws should guarantee equal access to DRAM/compute for all.

The Four Questions of AI

The piece lays out four coupled questions across the AI stack: how application companies build moats when models can absorb their features, who owns the harness and context layer given evidence that third-party harnesses outperform native ones, when foundation-model training costs stop paying back, and whether credit and power can keep pace with GPU demand. It cites figures including 98% input-token share, a 71% cost cut from smarter harnesses, roughly 83% of agent tasks routing to smaller models, $4t in projected five-year AI data center debt, and a B300 index at $12.29 per GPU-hour. It concludes that the layers behave like a double pendulum—simple individually but chaotic when coupled.

Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps

The authors propose an online reinforcement learning framework that trains UI navigation agents directly on live iOS applications, splitting the agent into a Planning agent and a UI Grounding model and using an LLM-as-a-Judge for dense step-level plus terminal rewards, optimized with PPO and step-level advantage estimation. They introduce GiGPPO, a neighbor-conditioned value estimation method that improves critic stability using reward signals from visually similar states across concurrent trajectories, backed by infrastructure orchestrating hundreds of concurrent virtual iOS devices. On UINavBench, iOSWorld, and AndroidWorld, their 8B RL agents substantially outperform SFT and terminal-only RL baselines, achieve nearly a sevenfold improvement over the SFT baseline in out-of-distribution generalization to Android, and outperform Gemini-3.5-Flash while tying Gemini-3.1-Pro on real iOS applications.

Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA

qwenGPU is an experimental MIT-licensed project by developer Vixhal that compiles a 4-lane SIMT GPU into 45 untrained layers placed in front of Qwen3.5-2B's own 24 layers, using the model's Gated DeltaNet state as registers, program ROM and global memory. The host compiles CUDA to PTX with clang and feeds it to the model as tokens, and the model writes each launch itself; the developer reports every launch was exact, a reference interpreter matched an NVIDIA Tesla T4 on 49 of 49 kernels, and text drift is a KL of about 7e-4. Reported speed is about 10 to 15 tokens per second on Apple MPS, and no prebuilt checkpoint is published.

How Small Can We Slice an AI Model?

The article argues that decentralized AI has been limited by pipeline parallelism, which assigns whole layers to individual machines and excludes devices with only a few gigabytes of spare RAM. It proposes sub-layer sharding instead: hosting individual experts from sparse mixture-of-experts models and compressed attention state on everyday devices, with a routing layer sending tokens only to the relevant shards. It surveys existing decentralized projects (Petals, Aethir, Akash, Exo) and frontier-lab work on sparse architectures (DeepSeek's MLA, Shazeer's research, DeepMind's Mixture-of-Depths), while noting the proposed swarm and memory budgets have not been demonstrated.

Hot take: AI skills getting obsolete after months

The post argues that several AI-related skills are becoming obsolete within months: context management is handled by automatic compaction in modern harnesses, prompt engineering is deprecated because models can recreate complex software from a one-liner, and sub-agent harnesses like Codex and Claude Code spawn sub-agents automatically without an extensive ecosystem. It also questions whether loops/goals are still necessary and whether Agents.md is a waste of tokens, citing https://arxiv.org/pdf/2602.11988. Two edits note a LinkedIn discussion at https://www.linkedin.com/feed/update/urn:li:activity:7517578946054447104/ and a comment by Andrej Karpathy that his prompts are getting shorter because over-specifying constrains the model.

Agentic AI is myth set in motion

Cosma Shalizi and the author argue in a new Economist piece that modern AI is best understood not as an emerging individual intelligence nor as mere parroting, but as a condensate of human culture—a self-mobilizing myth—since LLMs compress written knowledge (Barzun's "House of Intellect") into statistical summaries akin to the conventions learned by oral poets. Agentic AI combines these models with "harnesses" that let them issue instructions to ordinary programs and incorporate feedback, enabling feats like Lean-checked proofs or hacking attacks. Drawing on Shanahan et al.'s view of LLMs as non-deterministic simulators maintaining a superposition of simulacra, they describe agents as "many-pathed" like Odysseus—capable of extraordinary feats but without consistent goals.

Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon

Nace AI has released Drex Max v1, a dense 27B decision model with a 262,144-token window that returns option probabilities in one forward pass, as open weights under Nace.AI Open RAIL-M, shipping bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint and forks of llama.cpp and Ollama. It runs on NVIDIA GPUs, Apple Silicon Macs (24 GB+ for affine Q4, 48 GB for Q8_0) and CPU, and Nace reports 67.52 on the public Decision Index 0.4, ahead of Drex v2.0 at 61.25. Local Mac and CPU paths are text only for now.

dbreunig/nesformer

Inspired by Divyansh Agrawal's qwenputer, Drew Breunig built a working emulator for the Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B, so NES ROMs such as Super Mario Bros can run inside Qwen, very slowly. In a Hacker News comment he shared the Opus 5.5 prompts he used, which asked the model to explore emulating the chip and RAM with this technique and to build it for fun with weights and a CLI that can be pointed at a ROM. Opus reportedly found and used a collection of NES CPU conformance tests released 15 years ago by Christopher Pow.

AI being great at math makes human mathematicians more valuable, not less

The author argues that even if AI can solve extremely hard math problems, human mathematicians remain valuable because proofs must be understood and verified by humans, and expertise becomes scarcer and more valuable when AI makes it hard to distinguish correct from merely expert-looking content. Drawing on a field experiment with Süddeutsche Zeitung readers, the piece shows that exposure to AI-generated images lowered trust in news overall but increased demand for a trusted source, and applies the same logic to mathematics.

Samaya AI Releases Sterling: A 3B-Active Finance Research Agent Model that Beats Claude Fable 5.1 on FrontierFinance at 11x Lower Cost

Samaya AI released Samaya Sterling, a LoRA post-trained version of NVIDIA's Nemotron 3.5 Lightning (30B total, 3B active parameters) tuned as a financial research agent. On Samaya's self-reported FrontierFinance benchmark, Sterling scores 58.9% (60.6% in high-effort mode) versus 55.9% for Claude Fable 5.1, at $0.61 per question against $6.89. Training used LoRA on Tinker across two SFT stages and two GRPO stages, with a harness adapted for the small model.

Deterministic AI: Why LLMs Aren't, and How to Build Products That Are Anyway

This guide explains that LLMs are not deterministic in practice—a Thinking Machines test of 1,000 runs at temperature 0 on one prompt produced 80 different completions—because of provider batching, best-effort seeds, model alias updates and drifting inputs. It argues the deterministic-vs-probabilistic choice should be made per decision, keeping LLMs for reading and writing while moving money, access and compliance decisions into code rules, calibrated classifiers or decision models. It also offers layers for constraining LLM output and a method for measuring per-input agreement rates on borderline cases.

Baseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens

Baseten's Base Labs released joey-1, a 2B decision model post-trained from Qwen3.5-2B that returns a label and calibrated probability for choice, yes/no and score questions instead of generated text, priced at $0.01 per 1M input tokens with MIT-licensed weights on Hugging Face. Baseten reports 20 ms P50 latency for an average 2,000-token request and a score of 63.30 on the Jev Decision Index 0.4 public split, ranking #37 of 230 models and top under 4B, though it trails its 9B sibling elephant-1 by 8.94 points. It runs on Baseten Model APIs, dedicated deployments (262k context) or self-hosted on 1 CUDA GPU, and is also listed on OpenRouter.

Decisions for a penny: introducing joey-1, our fastest decision model

Base Labs introduces joey-1, a decision model with 3B or fewer parameters that returns a label and calibrated probability score in a single forward pass, serving an average 2,000-token request at a P50 end-to-end latency of 20 ms (about 2.5x faster than elephant-1) at $0.01 per million tokens. Post-trained from Qwen3.5-2B using the same data and recipe as elephant-1 and mammoth-1, it is available on Baseten Model APIs with open weights on Hugging Face, and supports a 16k context window on Model APIs versus the full 262k on dedicated deployments.

Speeding up language models for travel applications

Researchers combined knowledge distillation, which trains a compact student model on a large teacher's outputs, with Medusa, a framework that predicts multiple tokens in parallel, and applied the pipeline with Booking.com to three travel features. The final model matched the quality of a 70-billion-parameter teacher while running 10 to 20 times as fast, with reported gains including 2.9% higher engagement versus GPT-3.5 for AI Trip Planner and 15-times-faster Smart Filters responses for 99% of queries. The work received the Industry Track best-paper award at ACL.

Serving Prefill and Decode for Agentic AI (On Tenstorrent Blackhole)

Tenstorrent describes how its Galaxy systems, built on Blackhole chips, serve both prefill and decode phases on the same hardware so operators can shift capacity between them as agentic workloads change. The post explains how agent loops produce growing, largely cacheable context, models how faster decode only partly shortens dependent-call latency, and contrasts prefill's compute-bound long prompts with decode's memory-bound small batches. It cites a Wormhole FlashAttention kernel comparison (9 to 44 times faster than a basic implementation) as a measured example of data-placement control.

AI, Lean, and the Future of Formal Verification, with Lean Creator Leo de Moura

This is a link to a Data Exchange podcast episode featuring Leo de Moura, creator of Lean, discussing AI, Lean, and the future of formal verification. The item also links to its Hacker News discussion thread.

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral Large 4 ("Le Chonk") is a multimodal, multilingual Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context window, released in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License. It combines a 72-layer text decoder using MLA with a DeepSeek Sparse Attention indexer and a 1.6B Pixtral vision encoder, and the post reports 82% on the Artificial Analysis Cyber Index vulnerability reproduce-and-patch test and 93% on Cybench. Day-0 support covers Transformers, vLLM, llama.cpp, TRL fine-tuning, Inference Endpoints and Dell Enterprise Hub, with GGUFs from Unsloth and the llama.cpp team.

Day 0 support for Mistral Large 4 on vLLM

vLLM announces day-0 support for Mistral Large 4, a 1T-parameter MoE multimodal model with 49B active parameters per token, running on the V2 model runner with breakable CUDA graphs and a new vllm/models/mistral/ package. The work adds cross-layer KV sharing, DCP/PCP context parallelism, DSpark speculative decoding, NIXL-based disaggregated serving and HiSparse KV offload, plus day-0 ROCm support on AMD Instinct MI350 Series GPUs with MXFP4 experts. Support ships in the vllm/vllm-openai:mistral-large-4 Docker image, with FP8 and NVFP4 checkpoints and serving commands documented.

Le Chonk: Built for Coders, Trusted by Defenders.

Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal model trained from scratch in Europe, reporting leading open-weight results on coding benchmarks such as FrontierCode 2.0 (19.7%) and Terminal-Bench 4.0 (41%), plus 51% on the Artificial Analysis Cyber Index. It offers three reasoning modes (Low, Medium, High) and is priced at $0.70 input / $0.07 cached input / $2.10 output per million tokens under a modified MIT license that requires a separate commercial license for Model-as-a-Service offerings.

Build a better AI product with explicit classification

Dhruv Dang's newsletter explains AI classification, contrasting general-purpose models like GPT and Claude with purpose-built decision models such as Jev and Laya, and argues that explicit classification—making classification a separate step—gives teams granular control and flexibility over agentic workflows. A detailed example refactors an AI blogging agent so an intent-agent classifies each prompt into draft, save, publish, or unsupported requests, letting application code enforce rules like saving before publishing and route work to a sandboxed drafting-agent.

Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5

Google's Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v5 with a score of 59, leading 4 of 9 evaluations, and ranks second to Claude Opus 5.5 on the Coding Agent Index with 59 points. It scores 1609 Elo on AA-Briefcase (up 126 over Gemini 4 Argon) and 54 on AA-Omniscience v1.1, at $7.76 per Intelligence Index task on standard $4/$20 pricing. The model is being rolled out to selected users and is not yet publicly available, with a 50% launch discount and no confirmed end date.

Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design

Meta describes the next generation of its Andromeda ads retrieval engine, which uses a retrieval mixture-of-experts architecture and a single fused GPU pipeline to score every eligible ad at over one billion ads per second, a 100x throughput lift over the 2024 system. The system delivers near ranking-level model performance at retrieval scale with roughly 300x fewer inference-time FLOPs than the 2024 model, and Meta reports a 2.4% increase in conversions on Facebook since 2025. A joint collaboration with NVIDIA pushed individual components toward hardware limits, with the gating kernel reaching over 10 billion ads per second on a single GH200 and index matching exceeding 100 billion ads per second.

Polymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM Decoding

Researchers from Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute released FlashMap, a speculative decoding drafter that uses a continuous flow map to draft a full 16-token block in one forward pass, keeping tokens in the block consistent with each other. On Qwen3-8B the team reports up to 6.14x lossless speedup on GSM8K, and says 1-step FlashMap is the fastest drafter tested on every Qwen3-8B task at both temperatures 0 and 1. Drafter weights for Qwen3-8B and Qwen3-4B are on Hugging Face with vLLM and SGLang plugins.

Claude versus "GPU'less AI architecture" reasoning competition

The post reports a 1,186-question contest between SapienTronic AI (previously Minotaur AI) and Claude across a 69-part water-treatment plant scenario, bridged logic puzzles, and a trick question about rules only an endless town can obey. SapienTronic AI answered all 1,186 correctly in about 28 seconds on a laptop with no tokens or GPU, while Claude (Fable 5.1 and Opus 5.5) got 1,185 right, used about 528,041 tokens over roughly 28 minutes, and made one confidently wrong answer by treating a contradictory plant state as real. The author argues that a single confident mistake matters because it cannot be identified, can propagate through chains of decisions, and is not the kind of checkable evidence safety standards require.

Announcing Artificial Analysis Intelligence Index v5

Artificial Analysis released Intelligence Index v5, its biggest upgrade since v4, adding FrontierCode 2.0 and Terminal-Bench-Science and upgrading HLE-Diamond, AA-Omniscience v1.1 and CritPt v2, with 60% of the Index now based on private test sets or held-out answers. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55), while the open weights gap widens from 3 points to 12. It also launched the Progress Index v1, which tracks model capability over time on one continuous scale.

Running Local AI on Nvidia DGX Spark Without CUDA

CUDA++ is a research-stage toolchain and runtime that encodes GB10 (sm_121) machine code directly, replacing nvcc, PTX and libcuda with its own encoder and a client for NVIDIA's open kernel module ioctls, while NVIDIA's kernel module and signed GSP firmware remain. It currently supports only the DGX Spark's GB10, does not compile CUDA C++ or PTX, and reports partial coverage of CUDA's libraries; a head-to-head speed comparison against llama.cpp and vLLM has not been published.

Giotto Sovereign European AI Models

Giotto is a sovereign AI operating system that combines a reasoning model with an assistant, code, integrations, documents, agents and infrastructure monitoring, deployable on European cloud or private infrastructure. It offers two models, Giotto Alto and Giotto Solo, with benchmark comparisons against OpenAI, Anthropic, Google, Moonshot AI, Mistral, Nvidia and DeepSeek models across AIME 2026, GPQA-Diamond, HLE and Math-500. Giotto is available via its cloud, as a download for your own GPUs, or as a certified appliance with Giotto preinstalled.

AI refutes 30-year-old mathematical assumption in computer science

OpenAI claims its AI models have disproven the exponential time hypothesis, a nearly 30-year-old assumption underpinning much of computer science, by devising an algorithm that beats the previously assumed speed limit for solving logical puzzles. The result, announced on 26 October alongside claimed breakthroughs in mathematics and physics, was verified using the Lean proof-checking software and has been sent to an independent group including ETH's original proposers. Researchers say if confirmed it would require a complete revision of several areas of algorithms and complexity, though it does not directly resolve the P versus NP problem.

Featured coverage of this story: 《OpenAI says its AI disproved the Exponential Time Hypothesis》